Learning from oneself: Function learning with self-generated samples

Humans often learn functional relationships in environments shaped by their own prior actions rather than by repeated exposure to an objective ground truth. The present research investigated how people learn functions when learning is based on self-generated samples. Across three preregistered behavioral experiments (N ≈ 1,800), we introduced a novel experimental paradigm that occupies an intermediate position between experience-based learning and iterated learning. Participants repeatedly learned relationships between two variables across multiple rounds, where training data in later rounds were generated from their own prior predictions rather than from an externally defined function. The results revealed a characteristic combination of inductive bias, stability, and flexibility in human function learning. Learned relationships were systematically aligned with the true underlying functions, and the relative ordering of learned correlations in the final round preserved meaningful distinctions among functional forms. At the same time, learning trajectories exhibited strong path dependence: once formed, beliefs about functional structure tended to persist across rounds. Analyses of learning-phase behavior and cognitive modeling showed that differences across functions were primarily driven by initial error rather than learning rate. Together, these findings demonstrate that learning from self-generated samples does not inevitably lead to runaway bias. Instead, it supports adaptive knowledge acquisition while revealing how prior assumptions and feedback jointly shape learning over time. The present paradigm provides a new framework for studying learning in environments where experience is endogenously generated.